https://blog.yexca.net/en/Recent content on yexca'BlogHugo -- gohugo.ioenyexcaThu, 08 Oct 2026 00:00:00 +0900https://blog.yexca.net/en/archives/294/Thu, 08 Oct 2026 00:00:00 +0900https://blog.yexca.net/en/archives/294/As you buy more and more doujin voice works, your folders get longer and longer. Sometimes you only remember the cover art and the voice actor (CV) but can’t recall the title; other times, you fall asleep halfway through last night and have to drag the progress bar around the next day to find your spot. Even though the files are right there on your hard drive, finding them and resuming playback still takes some effort. Kikoto is an open-source doujin voice library I created …https://blog.yexca.net/en/archives/293/Fri, 11 Sep 2026 00:00:00 +0900https://blog.yexca.net/en/archives/293/A Systems View of Agentic Language Model Capability GPT-6-Astra Abstract Large language models are often compared as if their weights were the complete agent. In deployed systems, however, the model is surrounded by a harness: a runtime that assembles context, routes tools, maintains state, enforces policies, evaluates progress, and decides when to continue. We argue that this layer is not incidental glue but a first-class source of capability. We introduce a minimal …https://blog.yexca.net/en/archives/292/Mon, 24 Aug 2026 19:43:06 +0900https://blog.yexca.net/en/archives/292/Although I had seen this widely circulated quote before, I didn’t know its source at the time. Recently, by chance, I learned about LTT and how Linus used AI to fix Linux bugs, mentioning how difficult the process was and how much help he received from AI. Then I happened to come across this video. Although it’s a video from late 2025, the topics discussed in it almost completely dissolved my anxiety about AI and overturned some of my own thinking. This article records some …https://blog.yexca.net/en/archives/291/Wed, 19 Aug 2026 08:15:06 +0900https://blog.yexca.net/en/archives/291/This article introduces several AI image generation modes, specifically t2i, i2i, and r2i, which represent Text-to-Image, Image-to-Image, and Reference-to-Image, respectively. It also includes the i2t2i mode, which is a pipeline of Image-to-Text followed by Text-to-Image. Basic Concepts The core components of AI image generation (taking Stable Diffusion as an example) include: Base Model (UNet/DiT): Responsible for denoising and feature prediction in the low-dimensional …https://blog.yexca.net/en/archives/290/Thu, 13 Aug 2026 19:01:11 +0900https://blog.yexca.net/en/archives/290/Ever since I learned about the scp command, I have rarely used FTP services or other strange workarounds to transfer files from my local machine to a server. However, recently I needed to transfer a relatively large file. Because the transfer time was long and I was afraid of corruption due to interruption, I started looking for other solutions. During this process, the concept of pipelines reminded me of an operation I encountered when playing around with OpenWrt, so I …https://blog.yexca.net/en/archives/289/Mon, 06 Jul 2026 11:47:17 +0900https://blog.yexca.net/en/archives/289/In December 2024, I tried writing my very first Windows GUI software, Stop Looking at the Screen, Go walk (去散步). yexca/yasumiProject Loading... - - Although it was very simple, it felt like my first time developing without relying heavily on a framework. At the same time, I was determined to use Go no matter what back then. I happened to find a framework that fit the theme so well, creating a pun that echoed the theme with the tech stack. So looking back at these notes today …https://blog.yexca.net/en/archives/288/Sun, 28 Jun 2026 02:00:40 +0900https://blog.yexca.net/en/archives/288/When starting a service, you may occasionally encounter a situation where a port is occupied. However, when you actually query it using netstat -ano | findstr :[prot], you might find that no program is occupying it. This is usually because it has been occupied by Windows as a reserved port. If a program is indeed occupying it, you can use the taskkill command to terminate the process. Generally, simply changing to another port will solve the issue. However, when you want to …https://blog.yexca.net/en/archives/287/Mon, 22 Jun 2026 01:26:59 +0900https://blog.yexca.net/en/archives/287/Lately, when my local PC is training models, it completely locks me out of doing other work. Using my Mac or Linux devices feels a bit off too, since they already have their own dedicated tasks. So, I figured I’d put the lab computer to use. Once set up, I can run my training there and keep doing my own thing on my main machine. After doing some digging, I decided to set up three different connection methods to cover my bases: Microsoft’s official RDP, VS Code Remote …https://blog.yexca.net/en/archives/286/Tue, 16 Jun 2026 13:07:08 +0900https://blog.yexca.net/en/archives/286/I used to think locally deployed models could speak completely freely. But after hands-on experience, I realized modern AI models have two layers of shackles: one is the cloud provider’s external filter, and the other is the “refusal neurons” deeply baked into the model weights during safety alignment training. To get the perfect experience, you need to deploy an uncensored model. Warning Uncensored versions generally remove safety guardrails completely and will execute any …https://blog.yexca.net/en/archives/285/Sat, 13 Jun 2026 13:02:57 +0900https://blog.yexca.net/en/archives/285/Back when OpenClaw was super popular, I wasn’t really a fan of using Node for deployment. The subsequent security vulnerabilities confirmed my doubts. Now, there are plenty of great forks. Since I prefer lightweight setups, I went with the lighter ZeroClaw. Since I’m deploying my own AI assistant, I don’t want my data uploaded to the cloud. So, the goal this time is to build a fully local private assistant. Overview To keep everything fully local, I’ll use Ollama for hosting …https://blog.yexca.net/en/archives/284/Sun, 07 Jun 2026 18:42:10 +0900https://blog.yexca.net/en/archives/284/A friend strongly recommended Codex to me recently, so I gave it a spin. After trying it myself, I’d call it an “idea extractor.” Before this, my impression of AI coding was that it could build things, but the resulting code would be a mess to maintain. But after actually getting my hands dirty, I burned through 72M tokens in one go. My ideas kept getting pulled out and turned into reality. It was a massive productivity explosion. What I Built Let’s talk about what I actually …https://blog.yexca.net/en/archives/283/Tue, 02 Jun 2026 17:23:44 +0900https://blog.yexca.net/en/archives/283/Completely relying on AI to extract clean vocals is still not quite realistic with current technology; in most cases, manual correction is needed. However, as long as you have enough data, you can always scrape together enough AI-extracted vocals suitable for training. This article discusses a workflow for filtering training audio using only AI, from the perspective of an outsider. Everything used here is open-source. Training vs. Inference Audio My recent research on AI …https://blog.yexca.net/en/archives/282/Fri, 29 May 2026 08:58:06 +0900https://blog.yexca.net/en/archives/282/I recently checked my email and found Sparkle (a character in Honkai: Star Rail) quite interesting, so I hopped back into the game. (Yes, Sparkle. I actually forgot to publish this post after writing it, and as usual, I ended up uninstalling the game after playing for just a few days). Maybe because it’s been a while, but this playthrough felt different. I didn’t feel like a player; instead, I felt like I was trying to reverse-engineer and understand its system architecture. …https://blog.yexca.net/en/archives/281/Thu, 28 May 2026 15:52:18 +0900https://blog.yexca.net/en/archives/281/Set up intranet penetration using frp with a full Docker deployment. Tested the speed, and it’s pretty good with acceptable latency. DNS Configuration You’ll need a server with a public IP. I handle my SSL certificates via Cloudflare, so all my connections go through port 80. Configure a few domains pointing to your server: one for connecting to frp (e.g., frp.example.com), and others for the services you want to proxy (e.g., server1.example.com, server2.example.com). Note …https://blog.yexca.net/en/archives/280/Sat, 23 May 2026 03:00:06 +0900https://blog.yexca.net/en/archives/280/Getting Audio For training, real lossless audio is highly recommended (specifically pcm_f32le WAV files). It’s best to rip them from physical albums. You can refer to this guide: https://www.bilibili.com/opus/925630344961458181 , but skip the FLAC conversion and export directly to WAV. This guide uses Applio (based on the RVC architecture). You’ll need about 10 to 60 minutes of clean vocals (acapella) to get good training results. Don’t exceed 2 hours. Remember, quality over …https://blog.yexca.net/en/archives/279/Mon, 18 May 2026 20:53:11 +0900https://blog.yexca.net/en/archives/279/Figures Add the following code to your preamble. The second line sets the default image path to the img/ folder. 1 2 3 % Image \usepackage{graphicx} \graphicspath{{img/}} % set default path to img/ Standard usage: 1 2 3 4 5 6 \begin{figure}[htbp] \centering \includegraphics[width=0.8\textwidth]{example.jpg} \caption{caption} \label{fig:label} \end{figure} The htbp parameters stand for: h: here (current position) t: top b: bottom p: page (new page) The [width=0.8\textwidth] …https://blog.yexca.net/en/archives/278/Wed, 06 May 2026 20:05:32 +0900https://blog.yexca.net/en/archives/278/Installation and Environment Creation Official installation reference: Installing conda - conda documentation Download link: Download anaconda You can download Miniconda and set up environments manually as needed. Create a running environment: 1 conda create -n env_name python=3.13 -y Activate the environment: 1 conda activate env_name Installing PyTorch For the CPU version, use this command directly: 1 pip install torch torchvision torchaudio For the GPU version, visit the …https://blog.yexca.net/en/archives/277/Sun, 19 Apr 2026 22:20:06 +0900https://blog.yexca.net/en/archives/277/First, you need to install the go environment. 1 sudo apt install golang Then, initialize the module: 1 hugo mod init github.com/<your_user>/<your_project> Import the modules you want to add in your hugo.yml configuration file: 1 2 3 module: imports: - path: github.com/name/repo Update the configuration: 1 hugo mod get -u Module Storage Since Hugo uses Go dependencies, use this command to see where the modules are actually stored: 1 go env GOMODCACHE To vendor the …https://blog.yexca.net/en/archives/276/Wed, 15 Apr 2026 09:27:01 +0900https://blog.yexca.net/en/archives/276/When it comes to audio and video processing, FFmpeg is the ultimate open-source tool. I’ve been using it for a while, but my notes were scattered. Since I’m using it more frequently now, I decided to compile my most-used commands here (and add a few missing ones). Convert Video Format 1 ffmpeg -i video.mp4 video.avi Extract Audio from Video 1 ffmpeg -i input.mp4 -vn output.mp3 -vn: Disables video recording in the output file. Re-encode Video Sometimes a video file has a …https://blog.yexca.net/en/archives/275/Sun, 12 Apr 2026 22:21:30 +0900https://blog.yexca.net/en/archives/275/Defining Options API Options Vue 3.3 introduced the defineOptions() macro. It allows you to define Options API options (like name) without needing a separate script block, excluding props, emits, expose, and slots (which have their own dedicated define macros). 1 2 3 4 5 6 <script setup> defineOptions({ name: 'LoginIndex', // ... }) </script> Why use name? Debugging: The component name shows up in Vue DevTools. Recursive Components: Allows a component to call itself using its …